What Is an Authority Layer for AI Agents?

An authority layer is the part of an agent system that decides, and records, when a person's permission is borrowed and what the fleet is allowed to do with it afterwards. Agents will reach things they cannot do alone: a login on a system that has no service account, a payment that needs a signature, a rule that lives in somebody's head. The authority layer is where a person steps in with their own authority, once, and where that permission and procedure become something the fleet reuses within stated limits.

The three parts

1. Intervention. An agent asks for help with the page attached. A named person takes over the agent's live browser or desktop, does the thing, and marks it done. Typed text is never recorded.
2. Compilation. A model reads the recording and writes a procedure, not a transcript: numbered steps, a "done when" line, and an authority section naming what still requires a human and what the agent must never do. If the recording does not support a procedure, it refuses to write one.
3. Reuse, with a record. The next agent that hits the same wall on the same site is answered from the skill. Every request, intervention, approver and outcome is kept as an oversight record you can export.

Why "authority" and not "human in the loop"

Human in the loop describes a moment: a person is consulted. Authority describes what was transferred in that moment and on what terms. The difference shows the second time the wall appears. A human-in-the-loop system pages the person again. An authority layer has a procedure with a boundary and an audit trail, so the person is paged only for the part that genuinely needs them.

What it is not

• Not observability. Tracing tells you what the agent did; it has no notion of permission.
• Not RPA. A recording replays coordinates; a skill states intent and stops at its boundary.
• Not a CAPTCHA solver, a credential vault or a way to hide automation. A person completes those steps in the open, on the live session.

The boundary, concretely

# Skill: submit the weekly timesheet in hr.example
Use when: the timesheet page is open and the week is complete.
1. Open "Submit for approval".
2. Confirm the week shown matches the current week.
3. Choose "Submit".
Done when: the page shows "Awaiting manager approval".

AUTHORITY:
requires-human: approving the timesheet, editing hours already entered
never: submitting a week that is not the current one

The two lines under AUTHORITY are the point of the whole layer. They are written at compile time from what the person actually did and did not do, they are served to every agent with the skill, and they are what the oversight record checks against.

Questions people ask

Who needs an authority layer?

Any team running agents against systems that were built for people: payer portals, ERPs, supplier sites, internal tools with SSO. The first sign is a person being paged for the same thing every week.

Is it a compliance tool?

It produces the record compliance asks for, who allowed what and when, but its job is operational: getting agents through walls without widening what they may do on their own.

Where do the skills live?

With the customer who learned them, scoped to the site they were learned on. They are not shared across customers.

What happens when a skill is wrong?

Agents report outcomes; a skill whose failures outnumber its successes stops being offered, and an operator can retire one on the spot with a reason that goes in the record.